A learning-based residual feed-forward and a condition-aware safety margin control strategy were developed and validated for an electro-hydraulic variable valvetrain. The overall goal is to enable stable, accurate and flexible operation while avoiding fatal valve-piston collision. Timing accuracy in such systems is limited by condition-dependent actuation delays and velocities that drift with oil temperature, viscosity, pressure, supply voltage, and aging which can differ across actuators in the same system. In the proposed scheme, a geometric backbone (earliest safe intake valve opening (IVO) and latest safe exhaust valve closing (EVC) derived in the crank-angle domain from piston kinematics) is augmented by online identification of per-actuator, per-action residual delays. Residuals are learned under strict gating (steady operation, no clipping, no faults) and then fused with the baseline feed-forward component so that steady bias is transferred from the PI integral action to the maps without destabilizing the loop. Uncertainty is quantified online via a lightweight variability metric, and the additional safety margin is expanded or contracted accordingly. The method was implemented and validated on an engine dynamometer using a rapid prototype controller with crank-synchronous sensing and actuation. PI integral contributions were observed to decay toward zero as learned residuals converged, while tracking errors remained below 1° of crank angle. In transient operation, RMSE was reduced by 64% for valve closing and 62% for valve opening with residual learning compared to the system without the proposed compensation. Because residuals and variability are tracked for each actuator, sustained drift provides a diagnostic possibility for preventive maintenance. Although demonstrated on a valvetrain, the proposed approach is applicable to delay-dominated, aging-sensitive actuators requiring precise timing.
Özdemir et al. (Thu,) studied this question.